diff options
author | Joseph Greathouse <Joseph.Greathouse@amd.com> | 2019-09-18 14:49:57 -0500 |
---|---|---|
committer | Alex Deucher <alexander.deucher@amd.com> | 2020-04-28 16:20:30 -0400 |
commit | b8020b0304c8f44e5e29f0b1a04d31e0bf68d26a (patch) | |
tree | 0fa3c05128379f6c334fbfb646d8f45efc95c1c5 /drivers/gpu/drm/amd/amdkfd/kfd_priv.h | |
parent | 29633d0e204df1e051d9036e4f493f228ac19fb4 (diff) |
drm/amdkfd: Enable over-subscription with >1 GWS queue
The current GWS usage model will only allows a single GWS-enabled
process to be active on the GPU at once. This ensures that a
barrier-using kernel gets a known amount of GPU hardware, to
prevent deadlock due to inability to go beyond the GWS barrier.
The HWS watches how many GWS entries are assigned to each process,
and goes into over-subscription mode when two processes need more
than the 64 that are available. The current KFD method for working
with this is to allocate all 64 GWS entries to each GWS-capable
process.
When more than one GWS-enabled process is in the runlist, we must
make sure the runlist is in over-subscription mode, so that the
HWS gets a chained RUN_LIST packet and continues scheduling
kernels.
Signed-off-by: Joseph Greathouse <Joseph.Greathouse@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_priv.h')
-rw-r--r-- | drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 13 |
1 files changed, 13 insertions, 0 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 43b888b311c7..d48b33449267 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -411,6 +411,10 @@ enum KFD_QUEUE_PRIORITY { * @is_active: Defines if the queue is active or not. @is_active and * @is_evicted are protected by the DQM lock. * + * @is_gws: Defines if the queue has been updated to be GWS-capable or not. + * @is_gws should be protected by the DQM lock, since changing it can yield the + * possibility of updating DQM state on number of GWS queues. + * * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid * of the queue. * @@ -433,6 +437,7 @@ struct queue_properties { bool is_interop; bool is_evicted; bool is_active; + bool is_gws; /* Not relevant for user mode queues in cp scheduling */ unsigned int vmid; /* Relevant only for sdma queues*/ @@ -564,6 +569,14 @@ struct qcm_process_device { */ bool reset_wavefronts; + /* This flag tells us if this process has a GWS-capable + * queue that will be mapped into the runlist. It's + * possible to request a GWS BO, but not have the queue + * currently mapped, and this changes how the MAP_PROCESS + * PM4 packet is configured. + */ + bool mapped_gws_queue; + /* * All the memory management data should be here too */ |